The question arrives weekly now, from clients, from peers, from the group chat: which AI is best? And the honest answer is that it is the wrong question, asked in the wrong unit. Nobody who buys media for a living asks which channel is best; they ask what the job is worth and what the cheapest instrument is that reliably does it. Model selection is the same discipline wearing a different interface, and after a year of running four subscriptions side by side for real client work, I can tell you the selection math matters more than the brand loyalty.
Start with the terrain, because it has quietly standardized. The four major platforms all sell a serious tier at roughly twenty dollars a month, Claude Pro, ChatGPT Plus, Gemini AI Pro, Perplexity Pro, and premium tiers between one and two hundred. And within the past year, all of them converged on the same rationing design: compute-metered usage inside rolling windows, Claude's five-hour window, ChatGPT's three-hour fast-model window with a separate weekly cap on its reasoning mode, Gemini's five-hour compute meter as of this spring. I wrote about what that convergence means for extracting value in your AI subscription is worth $25,000 a month; this piece is about the decision that comes before extraction, which model you point at which problem.
The mechanics that make selection a real economic decision: within each platform, the models are not priced alike against your quota. On Claude, an Opus message draws roughly five to ten times the budget of a Sonnet message. On ChatGPT, the fast model is generous, around 160 messages per window, while the deep reasoning mode lives under a much tighter weekly allowance. Gemini's Deep Think sits behind its Ultra tiers. Run every task through the flagship reasoning model and you are the media buyer who insists on prime-time television for a garage sale: the instrument works, the economics are absurd, and you run out of budget three days before you run out of month.
Which model should you use? Answer three questions.
The same routing logic from this piece, applied to your task. Model data checked July 15, 2026.
| Platform | $20-tier default | Flagship | Reach for it when |
|---|---|---|---|
| Claude | Sonnet 5 | Opus 4.8 (Fable 5 where available) | Expert writing, strategy, sustained reasoning, agentic coding |
| ChatGPT | GPT-5.6 (Terra tier) | GPT-5.6 Sol | Broadest toolbox, hard STEM reasoning, voice |
| Gemini | Gemini 3.5 Flash | Gemini 3.1 Pro | Million-token context, multimodal, price |
| Perplexity | Pro search | Max, with Model Council | Anything that lives or dies on citations |
So classify the work the way you would classify placements, by what an error costs. Throughput work, first drafts, summaries, reformatting, cleanup, subject lines, is where fast models live: Sonnet-class, ChatGPT's instant tier, Gemini's standard model. Errors are cheap because you are reviewing everything anyway, and volume is the whole point. Judgment work, strategy, analysis, anything where a wrong answer costs real money or reputation, is what the reasoning models are for, and it is exactly where their five-to-ten-times budget premium is a bargain rather than a tax. The test I use, and the one I gave readers in the subscription piece, applies here: if the task deserves a $25,000 consultant, give it the consultant model and load it with context; if it deserves a $500 freelancer, the freelancer model is not a compromise, it is correct sizing. Research work, anything that lives or dies on current sources and citations, belongs to instruments built for retrieval: Perplexity's whole product, or the Deep Research modes the others meter by the day.
On the brand question, the differences are real and smaller than the discourse suggests. Claude remains the strongest sustained thinker over long documents and the best writer in the set. ChatGPT has the broadest toolbox. Gemini's million-token context swallows entire contracts and data rooms whole. Perplexity is the researcher, with receipts. But the gap between the tools is smaller than the gap between users, and a twenty-dollar subscription driven with context and intent beats a two-hundred-dollar one used like a search bar. The pattern rhymes with the pricing-model lesson from the three ways to pay for advertising: what you buy matters less than whether you are the kind of operator who can extract what the contract actually contains.
Which points at the only selection method that survives contact with reality: run the bake-off yourself, on your own recurring tasks. Take the five prompts your team actually runs weekly, the brief, the analysis, the rewrite, and run them through the candidates blind, then score outputs without knowing which model produced what. It is an afternoon of work, it costs nothing beyond subscriptions you were evaluating anyway, and it settles arguments the benchmark charts never will, because public benchmarks measure the average of everyone's tasks and you are not average, the same reason you do not need a data science team to run a holdout beats trusting anyone else's dashboard. My own bake-offs reshuffle every few months as the labs ship; the framework never changes.
The framework, then, in the space of a media plan's first line: inventory your tasks by error cost, route throughput work to fast models, reserve reasoning models for the questions whose answers are worth multiples of the subscription, hand research to the retrieval tools, and re-test quarterly because this market moves faster than any auction I have ever bought. The right model was never the smartest one available. It is the cheapest one that clears the bar for the task in front of you, with the expensive one held in reserve for the moments that justify it. That is not a technology insight. It is allocation, the oldest discipline in this business, pointed at the newest line item on it.